Deep neural networks (DNNs) are powerful black-box predictors that have achieved impressive performance on a wide variety of tasks. However, their accuracy comes at the cost intelligibility: it is usually unclear how they make decisions. This hinders applicability to high stakes decision-making domains such as healthcare. We propose Neural Additive Models (NAMs) which combine some expressivity DNNs with inherent intelligibility generalized additive models. NAMs learn linear combination each...

We study a security threat to batch reinforcement learning and control where the attacker aims poison learned policy. The victim is learner / controller which first estimates dynamics rewards from data set, then solves for optimal policy with respect estimates. can modify set slightly before happens, wants force into target chosen by attacker. present unified framework solving poisoning attacks, instantiate attack on two standard victims: tabular certainty equivalence in linear quadratic...

Training set bugs are flaws in the data that adversely affect machine learning. The training is usually too large for manual inspection, but one may have resources to verify a few trusted items. of items not by itself be adequate learning, so we propose an algorithm uses these identify and thus improves Specifically, our approach seeks smallest changes labels such model learned from this corrected predicts correctly. We flag whose changed as potential bugs, can checked veracity human...

Generalized additive models (GAMs) are favored in many regression and binary classification problems because they able to fit complex, nonlinear functions while still remaining interpretable. In the first part of this paper, we generalize a state-of-the-art GAM learning algorithm based on boosted trees multiclass setting, showing that outperforms existing algorithms sometimes matches performance full complexity such as gradient trees. second part, turn our attention interpretability GAMs...

In reward-poisoning attacks against reinforcement learning (RL), an attacker can perturb the environment reward $r_t$ into $r_t+δ_t$ at each step, with goal of forcing RL agent to learn a nefarious policy. We categorize such by infinity-norm constraint on $δ_t$: provide lower threshold below which attack is infeasible and certified be safe; we corresponding upper above feasible. Feasible further categorized as non-adaptive where $δ_t$ depends only $(s_t,a_t, s_{t+1})$, or adaptive agent's...

Extracellular matrix (ECM), as an important framework for tumor microenvironment, plays roles in many critical processes, including growth, invasion, immune suppression, and drug resistance. However, few biomarkers of ECM-related genes (ERGs) have been developed prognosis prediction clinical treatment bladder cancer (BC) patients. Bioinformatics analysis LC-MS/MS were used to screen differentially expressed ERGs BC. Multivariate Cox regression Lasso construct validate ERGs-based prognostic...

In this paper, we study reinforcement learning in Markov Decision Processes with Probabilistic Reward Machines (PRMs), a form of non-Markovian reward commonly found robotics tasks. We design an algorithm for PRMs that achieves regret bound Õ((HOAT)^(1/2) + H²O²A^(3/2) H(T)^(1/2)), where H is the time horizon, O number observations, A actions, and T steps. This result improves over best-known bound, Õ(H(OAT)^(1/2)), MDPs Deterministic (DRMs), special case PRMs. When ≥ H³O³A² OA H, our leads...

Abstract Background Basement membrane (BM) is an important component of the extracellular matrix, which plays role in growth and metastasis tumor cells. However, few biomarkers based on BM have been developed for prognostic assessment prediction immunotherapy bladder cancer (BLCA). Methods In this study, we used BLCA public database to explore relationship between BM-related genes (BMRGs) prognosis. A novel molecular typing was performed using consensus clustering. LASSO regression construct...

Training set bugs are flaws in the data that adversely affect machine learning. The training is usually too large for man- ual inspection, but one may have resources to verify a few trusted items. of items not by itself be adequate learning, so we propose an algorithm uses these identify and thus im- proves Specifically, our approach seeks smallest changes labels such model learned from this corrected predicts correctly. We flag whose changed as potential bugs, can checked veracity human...

We study data poisoning attacks in the online setting where training items arrive sequentially, and attacker may perturb current item to manipulate learning. Importantly, has no knowledge of future nor generating distribution. formulate attack as a stochastic optimal control problem, solve it with model predictive deep reinforcement also upper bound suboptimality suffered by for not knowing Experiments validate our approach near-optimal on both supervised unsupervised learning tasks.

Given a sequential learning algorithm and target model, machine teaching aims to find the shortest training sequence drive model. We present first principled way such sequences. Our key insight is formulate as time-optimal control problem. This allows us solve by leveraging theoretical computational tools developed over past 60 years in optimal community. Specifically, we study Pontryagin Maximum Principle, which yields necessary condition for optimality of sequence. analytic, structural,...

We study black-box reward poisoning attacks against reinforcement learning (RL), in which an adversary aims to manipulate the rewards mislead a sequence of RL agents with unknown algorithms learn nefarious policy environment priori. That is, our attack makes minimum assumptions on prior knowledge adversary: it has no initial or learner, and neither does observe learner's internal mechanism except for its performed actions. design novel attack, U2, that can provably achieve near-matching...

In E-commerce, a key challenge in text generation is to find good trade-off between word diversity and accuracy (relevance) order make generated appear more natural human-like. improve the relevance of results, conditional generators were developed that use input keywords or attributes produce corresponding text. Prior work, however, do not finely control automatically sentences. For example, it does put relevant ones first. Moreover, explicitly balance accuracy. To remedy these problems, we...

Theoretical researchers of manager psychology have excellent potential to extend its research framework more enterprise application areas, such as innovation, performance, and safety in production. Research these areas has also been increasing the past 10 years. Psychological capital is composed four aspects: self-efficacy, hope, optimism, tenacity. It plays an essential role stimulating organizational growth improving performance. In management work, managers, core members organization, a...

Large-scale sequencing plays important roles in revealing the genomic map of ccRCC and predicting prognosis therapeutic response to targeted drugs. However, relevant clinical data is still sparse Chinese population. Fresh tumor specimens were collected from 66 patients, then RNAs subjected whole transcriptome (WTS). We comprehensively analyzed frequently mutated genes our hospital's cohort as well TCGA-KIRC cohort. VHL gene most ccRCC. In cohort, BAP1 PTEN are significantly associated with a...

Abstract Background Adrenocortical carcinoma (ACC) is a rare endocrine neoplasm, which characterized by poor prognosis and high recurrence rate. Novel reliable prognostic metastatic biomarkers are lacking for ACC patients. This study aims at screening potential therapeutic targets of through bioinformatic methods immunohistochemical (IHC) analysis. Methods In the present study, using Gene Expression Omnibus (GEO) database we identified differentially expressed genes (DEGs) in validated these...